Comparing AI 3D Character Generators: Does the Humanoid Rig and Retarget?

What a rig-ready character needs

The mesh must be in a pose a rigger can use. Requirements differ by rigger, so check the one you will use; as a working assumption, hands should not be fused to the torso and legs should be separate. StraySpark reports that AccuRig needs A-pose meshes facing a specific direction and is scale-sensitive, which is the sort of input requirement that a generator’s “stylish” pose can silently violate. Check the pose before you spend rig credits. Generators can help at the source: Meshy’s web app has a Pose control (A-Pose or T-Pose), and its API has pose_mode.

Published pose guidance is not uniform:

Pick the pose your final skeleton expects, then generate in that pose, so you are not relying on a correction step later.

Rigging as documented

• Meshy web app lists Rigging at 0 credits. Its guide offers Humanoid, Quadruped Dog and Smart Rig (Beta) for other shapes. On humanoids Meshy places joint markers automatically and you drag each one onto its body part. Uploads can be up to 100 MB and need no textures; a generated model over 300,000 faces must be remeshed first. Smart Rig models “can’t use the animation library” and download as GLB only. Facial bones and tail physics are out of scope, per the guide.
• Meshy API lists Rigging at 5 credits and each Animation at 3. It “works well only with standard humanoid (bipedal) characters”, needs a textured model facing +Z, caps input at 380,000 triangles, and returns a 422 error when pose estimation fails. The web app and API are priced separately, and these limits differ between them, so use the figures for the route you will actually pay through.
• Tripo API lists Auto Rig at 25 credits, Rig Check free, and Animation Retarget at 10 credits per animation (up to five per request). Input is GLB, FBX, OBJ or STL up to 150 MB; output is GLB or FBX. Rig types are biped, quadruped, hexapod, octopod, avian, serpentine and aquatic, plus others on v3.0. The default model is still v1.0-20240301, biped only, and the recommended humanoid model v3.0-20260909 “must be requested explicitly”. A log that says only “Tripo auto-rig” is not reproducible.
• Rodin / Hyper3D added rigging API documentation in October 2026 per its changelog; its homepage does not mention rigging. The API spec lists a base charge of 0.5 credits, preset skeletons (human, quadruped, bird, dragon, kaiju), a custom hierarchy option, and stock animation IDs. For uploaded models, a reference image is required unless material is None, and “Embedded materials and textures are not preserved.” The pricing page’s plan table starts API access at Business.
• Tencent HY 3D Global (hosted Hunyuan): an earlier reading of its billing page recorded Auto Rigging at 10 credits per call. The page served a bot check to our last automated check, so this could not be reconfirmed. Check it before budgeting.
• 3D AI Studio FAQ lists Rigging at 15 to 20 credits with “biped, quadruped, and other creature types”, and 12 animation presets at 15 each.

Documentation tells you a rig is offered and what it costs. It does not tell you how well it binds a hand or a cloak.

What each engine asks of the skeleton

A rig only counts once it binds in the engine you ship in, and each engine documents a different gate.

Skeleton naming is now something some tools let you choose, which changes the comparison. Tripo’s spec option offers tripo, mixamo, actorcore, unity, unreal and vrm conventions for bipeds, the last four requiring v3.0. Meshy’s guide says Download Settings lets you choose a skeleton template for a humanoid, with Mixamo as the default. Log the choice, or you will be comparing naming conventions rather than rigs.

What the one detailed public rig comparison we read covers

StraySpark’s May 14, 2026 article describes running a humanoid character mesh through Tripo, Meshy, Cascadeur, AccuRig and Mixamo on a Blender 5.1 to Unreal 5.7 pipeline (FBX export, driven with Mover 2.0). The article says its protocol targets a “tri count ~30K, single material”; it does not report the tested mesh’s actual count. The author-reported times to a test animation are about 3 minutes for Tripo, 8 for Meshy (mostly retargeting, with non-standard spine and clavicle names needing manual bone remapping), 10 for AccuRig (including a scale fix) and 12 for Mixamo; Cascadeur is described as a pose and animation tool, not primarily an auto-rigger. It reports Tripo’s rig is biped-only with no facial joints and over-smoothed shoulder and chest weights, and that Meshy’s hip weighting was inconsistent, with hip pivots sometimes 5 to 15 cm off and causing foot sliding. AccuRig is reported to need A-pose meshes facing a specific direction and to be scale-sensitive.

Limits that matter to you: the article does not say how many characters it used (it mentions a non-humanoid stylized robot and a character that arrived at 0.01x scale); Unreal and Blender only, with no Godot or Unity retarget; it says it used “the same character mesh” but its protocol also allows “auto-generate-and-rig where the tool produces the mesh too”, so the generator and the rigger are not cleanly separated; hands are not mentioned and faces are covered only partly (jaw weights; facial rigs are described as still needing human work); no runs-per-tool or triangle counts are reported; and StraySpark sells paid tools for Unreal, Unity, Blender and Godot. So it says more about riggers than about generators. For the wider rigging picture see the rigging hub.

How to check your own characters

1. Generate C01 to C06 from the standard prompt set per generator; log plan and model version.
2. For each output, record pose gate result (limbs separated, hands visible), triangle count, and loose pieces using the script on the prompt-set page.
3. Rig it two ways: the generator’s own auto-rig (if any) and one shared tool for all, so you can separate the generator from the rigger.
4. Export FBX or GLB, import into your engine, bind to the standard skeleton your project uses, and apply the same walk, run and idle clips (retargeting notes).
5. Check poses with the arms raised, knees bent and the head turned. Add the harder set that Studio Liddell, an animation studio that sells rigging services, recommends documenting: deep bend, body twist, planted foot and prop interaction. Record tearing at shoulders, hips and fingers, foot sliding, and the minutes of fixing it took.

Fields added for characters: pose_gate (pass/fail), rig_source (own/shared), rig_model_version, skeleton_spec, bone_count, retarget_success (pass/fail), import_warnings, weight_defects (list, by region), minutes_to_walk_cycle.

Symptoms usually point to a layer of the pipeline. The first three rows come from Studio Liddell’s guide and the last three from Meshy’s rigging troubleshooting table, so both are vendor explanations to check against your file.

Pass or fail

Pass means the pose gate holds, the skeleton binds without manual bone remapping, the walk cycle plays with no visible tearing at shoulders or hips, and fixes take less time than you decided up front. The two rig runs tell you what to change:

Where the sources disagree or stop

The sources conflict in ways that change how you should read them.

• What Tripo rigs. StraySpark’s May 2026 article calls the rig “biped-only” and, in the same section, says Tripo handled a non-humanoid stylized robot better than any other tool. Tripo’s current docs list seven rig types, with a non-humanoid snapshot dated v2.5-20260210, before the article. Our inference, which the article cannot confirm, is that a default call would land on the biped-only v1.0 model. Either way, the article’s “biped-only” does not describe what the API documents now.
• Meshy’s bone names. StraySpark reports non-standard spine and clavicle names that needed manual remapping in Unreal. Meshy’s rigging guide now describes a skeleton template choice with Mixamo as the default for humanoids. The article may predate that option or may have used a different export path. Test the current download before planning around either account.
• How much cleanup to expect. Sorceress says its own humanoid auto-rig is “reliable enough to skip manual cleanup in most cases”. Studio Liddell, which sells rigging work, says “Treat an AI rig as an inspection candidate, never as a delivery asset.” Both have a commercial stake in their answer. Neither publishes rates.
• What the training data assumes. The HumanRig research dataset uses 11,434 T-posed meshes on one skeleton topology. That is one academic dataset, not a statement about commercial riggers, but it is a reason to treat unusual poses and proportions as the harder case.

Faces remain the agreed gap. Meshy’s guide sends facial bones to manual rigging, its comparison post says none of the three tools it covers produces a face rig that deforms properly out of the box, and StraySpark describes facial rigs as still needing human work.

Mixamo is covered in using Mixamo with AI-generated characters; non-humanoids need different tools, see non-humanoid and quadruped rigs. For characters starting from concept art, see from concept art to a 3D character.

Back to the comparison hub.